What is PlainSpeak?
PlainSpeak finds the phrases that make writing read as machine-made, then rewrites them in plain words while keeping your facts.
You paste in a draft. The scan marks the giveaways a reader notices, such as stock AI phrases, corporate speak and connector words like "Furthermore". When you rewrite, your AI provider rewrites the copy, and PlainSpeak checks that the numbers, dates, names, links and quotes in your original survived.
What you can do
- Scan a draft and see each tell highlighted, with a score out of 100.
- Rewrite it in a voice you choose: a light polish, a full rewrite, or a new post built from the same material.
- Check the facts in the rewrite against your original.
- Ask two other AI models to read the rewrite and point at anything that still sounds generated. This is the audit.
- Edit the rewrite by hand, with the scan updating as you type, a toolbar to format it, a Preview to see it formatted, and your original beside it in a wide window.
- Take it out formatted: Copy pastes with its headings, bold, bullets and links into Google Docs, Word and email, and Download .docx saves a Word file. Open a Word file to start from, too.
Why the tells matter
People are not good at spotting AI writing, but they act on what they suspect. In a 2026 study, 89 expert peer reviewers told human abstracts from ChatGPT ones only 52.5% of the time, and were just as confident when they were wrong (Richter and Larsson, *Research Methods in Applied Linguistics*). In another, people agreed less with arguments they believed an AI had written, whoever had actually written them (Garcia, Qian and Palminteri, *PLOS One*). Both found readers lean on surface signs: stock phrasing, formulaic structure, connector words and prose that is a little too polished.
So the wording matters whether or not anyone can really tell, and it matters for writing you did yourself, because a human draft that reads formulaic draws the same suspicion. On our test set of 18 machine drafts, a Rewrite with the default model cut the tells the scan counts from 10.2 to 2.2 per 1,000 words and kept every checkable fact. A plain rewrite by the same model, asked to keep the facts, left the tells at 10.9, dropped two facts and invented two, though two blind judge models rated it as human or more so. The phrase list is an editing heuristic, and no reader study of it has been run. The aim is copy that reads as yours, never a way to hide how it was written.
What PlainSpeak is not
PlainSpeak is a finishing tool, not an AI detector. The score counts wording a reader is likely to notice. It makes no claim about what a detection service would say about your text, and a clean score is not a promise that nobody will suspect AI.
What you need
| To do this | You need |
|---|---|
| Scan a draft | Nothing. Scanning is free and works offline. |
| Rewrite, check facts and find sources | A key from an AI provider, and either a free trial you have started or a licence. In the desktop app the 7 days start when you select Start the trial, not when you install it, and nothing that uses AI runs before that. More |
| Run the audit or Find sources | A key from OpenRouter in particular |
PlainSpeak runs on Windows and Mac computers. Your key is stored on your computer and sent only to the provider you chose. Exactly what PlainSpeak sends, and where.
For builders
- Bring your own key: OpenRouter, Anthropic, OpenAI, or any OpenAI-compatible endpoint. See Choose a model.
- Pick the model for each rewrite. Draft in one, rewrite in another.
- Voices, templates and samples of your own writing steer the register. See Edit and add voices and Calibrate to your writing.
- The licence is an Ed25519-signed key, verified on your computer, with no sign-in.
Next steps
- Get started: install PlainSpeak and run your first scan and rewrite.
- Find your way around: what each part of the window does.